Where to host computer-use agents (CUA)

A computer-use agent works the way a person does: it looks at a screen, moves a mouse, types, clicks, and reads what happens next. Models like OpenAI's CUA and Anthropic's computer use provide the brain, but the brain needs a body: a real desktop with a display, a browser, and applications it can actually operate. The question every team hits is where that desktop should run.

What a computer-use agent needs

To drive a computer, the agent needs a graphical desktop, not just a shell. It needs a rendered screen it can see, a browser it can log into, a mouse and keyboard it can move, and a file system that survives between steps. For real work it also needs to keep state: stay logged in, keep its downloads, and pick up where it left off tomorrow. And you need to watch it, because a computer-use agent that clicks the wrong button should be observable, not a black box.

The options

Your own laptop. Simplest to start: run the agent on the machine in front of you. But it uses your computer, works only while you are logged in, stops when the lid closes, and does not scale past one. Fine for a demo, wrong for production.

Self-hosted VMs. Spin up your own virtual machines with a desktop environment. You get full control, and you also get the full job: provisioning, isolation between agents, uptime, updates, and streaming the screen somewhere you can see it. One desktop is easy. A fleet of them, one per agent, with real isolation, is an infrastructure project.

Ephemeral browser or sandbox services. Cloud services that hand the agent a browser or a short-lived sandbox. Great for a single scrape or a quick code run, but often browser-only, with no persistence and a session timeout. A computer-use agent that needs desktop apps, a persistent login, or a multi-step job over hours outgrows them fast.

Managed persistent cloud desktops. A dedicated cloud desktop per agent, fully managed, always on. This is the shape that matches what a computer-use agent actually needs: a real screen, a real browser, persistence, isolation, and a way to watch it, without you running the infrastructure.

Why a persistent cloud desktop fits

Le Bureau gives each agent its own dedicated cloud desktop with a real browser, terminal, editor, and file system, streamed to your browser so you can watch the agent work in real time. Because it is a cloud machine, it stays on when your laptop is off and a long job finishes on its own. Each agent runs in its own isolated virtual machine, so one agent's mistake cannot reach another. Mission Control gives you one place to manage the whole fleet, and hosting is in Europe with data resident in the EU. You point your computer-use agent at it, bring your own AI key or use a managed model, and let it operate a real desktop instead of borrowing yours.

Related reading

• Le Bureau vs Claude Cowork: cloud desktop vs running on your own Mac.
• E2B alternatives for a full desktop: when a sandbox is not enough.
• Cloud desktops for AI agents: the full guide.

Getting started

You can launch your first cloud desktop on Le Bureau for free in about two minutes, connect your computer-use agent, and watch it click through a real browser and desktop. It keeps its state between sessions, so the agent that logs in today is still logged in tomorrow.